Pre-trained face recognition, minus the Caffe translation layer

What it does

The library provides Keras implementations of the Oxford VGGFace models—specifically vgg16, resnet50, and senet50 variants—converted from the original Caffe networks. It downloads pre-trained weights on first use and exposes the standard Keras Functional API for prediction, feature extraction, and fine-tuning.

The interesting bit

The real value is the grunt work it saves: someone already endured the weight conversion and layer mapping from Caffe to TensorFlow so you don’t have to. It also exposes the usual Keras levers like include_top=False and custom pooling, making it behave like any other application model in the ecosystem.

Key highlights

• Ships three architectures: vgg16 (2015), plus resnet50 and senet50 (2017).
• Supports feature extraction by dropping the classification head and specifying pooling.
• Includes preprocessing utilities (utils.preprocess_input) matched to each model version.
• Weights are cached locally after the first download.
• Code is MIT-licensed, though the underlying model weights carry Oxford’s original license.

Caveats

• The README pins compatibility to Keras 2.2.4 and TensorFlow 1.14.0, so expect friction with modern stacks.
• Theano backend is explicitly unsupported and untested.

Verdict

Worth a look if you need Oxford’s VGGFace weights in a Keras 2.x / TensorFlow 1.x environment. If your stack is newer, this repository may feel dated.